Page 58 - Contributed Paper Session (CPS) - Volume 7

CPS2028 Ayon M.
existing control. The next is the positive treatment effect, which refers to the
hypothetical experimental scenario where treatment A is more effective than
treatment B, or the new treatment performs better than the control. The third
scenario focuses on the negative treatment effect, which refers to the
hypothetical experimental scenario where treatment B is more effective than
treatment A, or, in the case of comparing a new treatment with a control, this
means that the new treatment is not as effective as the control. The procedure
used here is a fully sequential one that recalculates the randomization
probabilities for every new patient over 5,000 simulation runs for 400 patients,
each arriving sequentially into the trial. The censoring scheme considered here
is generalized type I right-censoring. The patient’s arrival pattern is simulated
from a uniform(0,365) distribution, whereas the response of a patient is added
to the recruitment time of the patient, and patients whose outcomes have not
been observed by a specified time are said to be generalized type I right-
censored. The length of the recruitment period is 365 days and the overall trial
duration is taken to be 581.66 days. Following Rosenberger, Vidyashankar and
Agarwal (2001), a covariate structure of three independent covariates have
been generated. These are Gender (Bernoulli, p = 0.5), Age (Uniform[30,75])
and Cholestrol Level (Normal [200,400]). The survival time of a patient with

covariate vector = (1, , , ) in treatment group is simulated from the
1
2
3

Weibull distribution with scale parameter () = exp( ) and shape

parameter = 1.07527. Since there are three predictive covariates in the

model, the direction and magnitude of the treatment difference will vary for
the patients, depending on their observed covariate values.
In survival trials, the delay time for a patient is the patient’s survival or
censoring time. To facilitate CARA designs with delayed responses, it is
required that, at the patient’s randomization time, only data from those
ℎ
patients who have responded before the patient’s arrival are used for
ℎ
computing the randomization probability for the patient. In practice, an
ℎ
assumption of immediate responses is infeasible due to inherent delay in time-
to-event outcomes. For the implementation of the CARA designs, initially 2
0
patients have been equally allocated to the two treatment arms using a Efron’s
Biased Coin design. Here, is a positive integer, and, following Sverdlov,
0
Rosenberger and Ryeznik (2013), it is chosen to be 40 for each treatment arm.
For appropriate implementation of the ERADE designs Hu, Zhang and He
(2009) recomended that it is reasonable to choose , the degree of
randomness of the design, to be between 0.4 to 0.7. When is smaller, the
ERADE is more deterministic and has a smaller variability. Hu, Zhang and He
(2009) showed that the ERADE response-adaptive designs give similar results
when = 0.5 as compared to when = 0.67. Here, for the implementation
of the ERADE designs, is chosen to be 0.55, whereas for the implementation

45 | I S I W S C 2 0 1 9